Immersive 3D worlds from any source
World Generator
The World Generator converts a text prompt, image, multi-image set, panorama, or video clip into a fully explorable immersive 3D world. Worlds are rendered using World Labs' Marble AI and delivered as a real-time gaussian splat scene you can walk through directly in the browser.
01
Create or open a session
Go to World Generator from the home page and click New Session, or open an existing session from the list. Each session remembers its settings and generation history.
02
Choose an input mode
Select how you want to describe your world: Text, Image, Multi-image, Panorama, or Video. Upload source files where required, then add an optional text prompt to guide the look and mood.
03
Configure and generate
Pick a model (marble-1.1 is the default), optionally set a seed for reproducibility, then click Generate World. Generation takes 3–6 minutes and progresses through three stages: Panorama → Draft → World.
04
Explore and export
Once complete, the 3D world loads in the interactive viewer. Move with WASD or arrow keys, drag with the mouse to orbit or pan, and scroll to zoom, then download the output panorama from the controls panel.
Input modes
Controls reference
Model
Choose between marble-1.1 (faster, standard quality-default) and marble-1.1-plus (slower, highest quality). Current model and variant costs are shown in the live table under Credits & pricing.
Input Mode
Selects the source type: Text, Image, Multi-image, Panorama, or Video. Switching input mode resets the uploaded source files.
Source
The file upload area for image, multi-image, panorama, or video inputs. Drag and drop or click to browse. For Multi-image mode, assign a direction (Front / Back / Left / Right / Auto) to each image.
Text
An optional free-text prompt to guide the world's mood, atmosphere, and details. You can use this alongside image inputs-for example, upload a room photo and add "at dusk with warm candlelight".
Seed
An optional integer (0 – 4,294,967,295). Setting the same seed with the same inputs reproduces an identical world. Leave blank for a random result each time.
Disable recaption
When enabled, the AI does not re-describe your source image before generating the world. Use this when your text prompt already fully describes the scene and you don't want the model to override it with its own interpretation.
Generation stages
Panorama
~1 min
A 360° equirectangular panorama is generated from your source. This becomes the environmental reference for the 3D geometry.
Draft
~2 min
A low-resolution draft world is built from the panorama. The spatial structure, depth, and primary geometry are established.
World
~1–2 min
The final gaussian splat scene is generated and optimised. The viewer loads once this stage completes.
Output & viewer
3D viewer
An interactive gaussian splat renderer loads your completed world with Blender-style navigation. Move-WASD or arrow keys, hold Shift to move faster, E/Q to move up and down. Orbit-drag with the middle mouse button. Pan-drag with the right mouse button. Zoom-scroll wheel. The viewer runs in real-time in the browser with no plugin required.
Mobile controls
On touch devices: single finger drag to orbit and look around, pinch to move forward or backward, and use the on-screen D-pad (bottom-left of the viewer) to strafe and move vertically. All D-pad buttons respond to touch hold.
Output panorama
The 360° equirectangular panorama generated in Stage 1 is shown in the controls panel. Click it to open a fullscreen lightbox, or use the Download button to save it as a JPG.
World description
If the model generates a caption for your world, it appears below the panorama. This is the AI's interpretation of the scene-useful as a prompt reference for future generations.
Session title
Click the session name in the nav bar to rename it. Titles are saved automatically and appear in your session history on the home page.
On-screen D-pad
A directional pad (↑ ← ↓ →) appears at the bottom-left of the viewer once a world has loaded. Hold any button to move continuously. E moves up and Q moves down. Works with touch on tablets.
Viewer keyboard shortcuts
Camera movement
View controls
Credit cost: World generation costs vary by model and input mode; see the live table under Credits & pricing. If generation fails on the World Labs side, credits are automatically refunded.
Use panoramas for best fidelity
A high-resolution equirectangular panorama (2560 × 1280 px or larger) gives the model the richest spatial information and typically produces the most accurate world geometry.
Add text alongside images
The Text field works in conjunction with any image input. Use it to specify mood, time of day, or details that aren't visible in the source-e.g. 'twilight, warm light through windows'.
Use seeds for iteration
Set a seed to lock the randomness and explore only prompt or input changes. Without a seed, the same inputs will produce slightly different worlds each time.
marble-1.1 for speed, plus for final
Use marble-1.1 to iterate quickly and validate your concept. Switch to marble-1.1-plus for the final generation when you're happy with the direction.